TRUST: A Framework for Decentralized AI Service v.0.1

The TRUST framework, introduced by researchers Yu-Chao Huang and colleagues, addresses critical vulnerabilities in centralized AI auditing. Large Reasoning Models (LRMs) and Multi-Agent Systems (MAS) in high-stakes domains suffer from single points of failure, scalability bottlenecks, opaque auditing, and privacy risks. TRUST (Transparent, Robust, and Unified Services for Trustworthy AI) tackles these with three innovations: Hierarchical Directed Acyclic Graphs (HDAGs) that decompose Chain-of-Thought reasoning into five abstraction levels for parallel distributed auditing; the DAAN protocol that projects multi-agent interactions into Causal Interaction Graphs (CIGs) for deterministic root-cause attribution; and a multi-tier consensus mechanism among computational checkers, LLM evaluators, and human experts with stake-weighted voting. This design guarantees correctness under up to 30% adversarial participation, as proven by a Safety-Profitability Theorem that ensures honest auditors profit while malicious actors incur losses. All decisions are recorded on-chain with privacy-by-design segmentation to prevent reconstruction of proprietary logic.

In benchmarks across multiple LLMs, TRUST achieved 72.4% accuracy—4-18% higher than baselines—and remained resilient against up to 20% corruption. The DAAN protocol reached 70% root-cause attribution (vs. 54-63% for standard methods) with 60% token savings, making it both more accurate and more efficient. Human validation confirmed the design's reliability (F1=0.89, Brier=0.074). TRUST enables four critical applications: decentralized auditing, tamper-proof leaderboards, trustless data annotation, and governed autonomous agents. This work represents a major step toward safe, accountable deployment of reasoning-capable AI systems, offering a practical blueprint for transparent and decentralized governance in an age of increasingly autonomous models.